Targs = [tensor<32xf32>, tensor<64xf32>, tensor<8xf32>, tensor<2xf32>],
      Tresults = [tensor<32xf32>, tensor<64xf32>, tensor<8xf32>]}
      : (tensor<32xf32>, tensor<64xf32>, tensor<8xf32>, tensor<2xf32>, tensor<2x!tf_type.string>)
        -> (tensor<32xf32>, tensor<64xf32>, tensor<8xf32>)
    tf_device.return %0#0, %0#1, %0#2 : tensor<32xf32>, tensor<64xf32>, tensor<8xf32>

  %max = arith.constant dense<[14.0, 13.0, 8.0, 0.0]> : tensor<4xf32>
  %mini = "tf.Identity"(%min) : (tensor<4xf32>) -> tensor<4xf32>
  %maxi = "tf.Identity"(%max) : (tensor<4xf32>) -> tensor<4xf32>
  %fq = "tf.FakeQuantWithMinMaxVarsPerChannel"(%in, %mini, %maxi) {num_bits = 4, narrow_range = true} : (tensor<3x3x3x4xf32>, tensor<4xf32>, tensor<4xf32>) -> tensor<3x3x3x4xf32>

// CHECK-NEXT:  %[[quant:.*]] = "tfl.quantize"(%[[split]]#0) <{qtype = tensor<2x!quant.uniform<u8:f32, 1.000000e+00>>}> {volatile} : (tensor<2xf32>) -> tensor<2x!quant.uniform<u8:f32, 1.000000e+00>>
// CHECK-NEXT:  return %[[quant]] : tensor<2x!quant.uniform<u8:f32, 1.000000e+00>>

    %cst_0 = "tf.Const"() {value = dense<1.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
    %var_0 = "tf.VarHandleOp"() {shared_name = "var_0"} : () -> tensor<!tf_type.resource<tensor<2xf32>>>
    "tf.AssignVariableOp"(%var_0, %cst_0) : (tensor<!tf_type.resource<tensor<2xf32>>>, tensor<2xf32>) -> ()
    %cst_1 = "tf.Const"() {value = dense<2.000000e+00> : tensor<3xf32>} : () -> tensor<3xf32>

    "tfl.yield"(%2) : (tensor<8xf32>) -> ()
  }) {num_bits = 5, narrow_range = false} :  (tensor<8xf32>, tensor<f32>, tensor<f32>) -> tensor<8xf32>
  func.return %rst : tensor<8xf32>

// CHECK: %[[CONSTANT:.*]] = arith.constant dense<0.000000e+00> : tensor<8xf32>
// CHECK: %[[QUANTIZE:.*]] = "tfl.quantize"(%[[CONSTANT]]) <{qtype = tensor<8x!quant.uniform<u8:f32, 1.000000e+00>>}>

// CHECK: return %[[DEQUANTIZE]] : tensor<8xf32>
}

// CHECK-LABEL: fakeQuantNotFolded
func.func @fakeQuantNotFolded(tensor<8xf32>, tensor<f32>, tensor<f32>) -> (tensor<8xf32>) {
^bb0(%arg0: tensor<8xf32>, %arg3: tensor<f32>, %arg4: tensor<f32>):
  %1 = "tf.FakeQuantWithMinMaxVars"(%arg0, %arg3, %arg4) {num_bits = 5, narrow_range = false} : (tensor<8xf32>, tensor<f32>, tensor<f32>) -> tensor<8xf32>
  func.return %1 : tensor<8xf32>

  %2 = "stablehlo.concatenate"(%0, %1) {dimension = 0 : i64} : (tensor<?xf32>, tensor<?xf32>) -> tensor<?xf32>
  %3 = "quantfork.stats"(%2) {layerStats = dense<[-0.83811146, 2.4960899]> : tensor<2xf32>} : (tensor<?xf32>) -> tensor<?xf32>
  %4 = "quantfork.stats"(%arg2) {layerStats = dense<[-1.5726943, 1.07351148]> : tensor<2xf32>} : (tensor<?xf32>) -> tensor<?xf32>

  func.return %1, %2 : tensor<*xf32>, tensor<*xf32>

// CHECK:  %0 = "tf.Gather"(%arg0, %arg2) <{validate_indices = true}> : (tensor<*xf32>, tensor<i32>) -> tensor<*xf32>
// CHECK: return %0, %arg0 : tensor<*xf32>, tensor<*xf32>
}

// -----

// CHECK-LABEL: tensorlistStackWithConstantElementShape
func.func @tensorlistStackWithConstantElementShape(%arg0: tensor<?x3xf32>) -> (tensor<2x3xf32>) {

      %a, %control1 = tf_executor.island wraps "tf.A"() {_class = ["loc:@class"]} : () -> (tensor<2xf32>)
      %s:2, %control2 = tf_executor.island wraps "tf.Split"(%c, %a) {num_split = 2 : i32} : (tensor<i32>, tensor<2xf32>) -> (tensor<1xf32>, tensor<1xf32>)
      tf_executor.fetch
    }
    func.return
  }

    "tf.AssignVariableOp"(%2, %arg0) : (tensor<!tf_type.resource<tensor<?xf32>>>, tensor<?xf32>) -> ()
    return
  }
}

// -----

// CHECK-LABEL: module
module attributes {tf_saved_model.semantics} {
  // CHECK: ml_program.global{{.*}}[[V:@.+]]({{.*}}) : tensor<3xf32>
  // CHECK: func.func @read_inserts_cast_if_necessary(%arg0: tensor<?xf32>
  // CHECK: [[V1:%.*]] = ml_program.global_load